USER: Unsupervised Structural Entropy-Based Robust Graph Neural Network

Unsupervised/self-supervised graph neural networks (GNN) are susceptible to the inherent randomness in the input graph data, which adversely affects the model's performance in downstream tasks. In this paper, we propose USER, an unsupervised and robust version of GNN based on structural entropy, to alleviate the interference of graph perturbations and learn appropriate representations of nodes without label information. To mitigate the effects of undesirable perturbations, we analyze the property of intrinsic connectivity and define the intrinsic connectivity graph. We also identify the rank of the adjacency matrix as a crucial factor in revealing a graph that provides the same embeddings as the intrinsic connectivity graph. To capture such a graph, we introduce structural entropy in the objective function. Extensive experiments conducted on clustering and link prediction tasks under random-perturbation and meta-attack over three datasets show that USER outperforms benchmarks and is robust to heavier perturbations.

Authors

• Yifei Wang South China University of Technology
• Yupan Wang The University of Auckland
• Zeyu Zhang Huazhong Agricultural University
• Song Yang
• Kaiqi Zhao The University of Auckland
• Jiamou Liu The University of Auckland

Keywords

• ML: Graph-based Machine Learning
• ML: Unsupervised & Self-Supervised Learning